Domain Adaptation of Deformable Part-Based Models.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object classifier accuracy degrades due to domain shift between training and application data.
• Adapting classifiers to specific scenarios is crucial for reliable performance.

Purpose of the Study:

• To develop novel domain adaptation (DA) methods for object detection, specifically for pedestrian detection using the Deformable Part-based Model (DPM).
• To create methods that adapt pre-learned classifiers using limited target-domain data without revisiting source data.

Main Methods:

• Introduced adaptive structural SVM (A-SSVM) for classifier adaptation between domains.
• Proposed structure-aware A-SSVM (SA-SSVM) incorporating feature space structure (e.g., DPM parts).
• Developed self-adaptive DPM using self-paced learning (SPL) and Gaussian Process Regression (GPR) for scenarios lacking annotated target data.

Main Results:

• The proposed A-SSVM and SA-SSVM methods effectively adapt DPMs for pedestrian detection.
• Adaptation was tested from synthetic/general datasets (PASCAL VOC) to on-board camera data.
• The methods successfully mitigated accuracy drops, preventing losses as high as 15 points compared to non-adapted detectors.

Conclusions:

• Novel domain adaptation techniques significantly enhance object detection performance in varied scenarios.
• The proposed methods offer effective solutions for domain shift problems in object classification, particularly for pedestrian detection.
• The developed approaches are robust, requiring minimal annotated target data and even functioning without it via self-adaptive strategies.